> Paired with the fact it's not much faster in practice than epoll in a large majority of usecases
It just takes significant development effort to take advantage of io_uring's unique ability to coalesce many SQEs into a single syscall. I'd argue a minority of use cases involve a single syscall per logical work slice, not the majority. But most programs aren't yet written to exploit it, not for lack of potential benefit.
There's similarity to SIMD extensions in this regard, but wide batching of syscalls is arguably far more generally applicable than SIMD instructions.
When you make a low-effort conversion of an epoll-utilizing program to io_uring, it's similar to plugging some SIMD intrinsics into your existing program without actually refactoring anything to profitably go wide and avoid continuously converting to/from the wide types. You'll find it's either no faster or even slower than before, but that doesn't mean SIMD can't make the application faster, it just takes some proper doing.
It took years for SIMD to become well utilized once C compilers added intrinsics exposing them to programmers. I expect a similar delay before userspace evolves to exploit io_uring, and we could really use better language-level constructs to make writing such async code more ergonomic.
It just takes significant development effort to take advantage of io_uring's unique ability to coalesce many SQEs into a single syscall. I'd argue a minority of use cases involve a single syscall per logical work slice, not the majority. But most programs aren't yet written to exploit it, not for lack of potential benefit.
There's similarity to SIMD extensions in this regard, but wide batching of syscalls is arguably far more generally applicable than SIMD instructions.
When you make a low-effort conversion of an epoll-utilizing program to io_uring, it's similar to plugging some SIMD intrinsics into your existing program without actually refactoring anything to profitably go wide and avoid continuously converting to/from the wide types. You'll find it's either no faster or even slower than before, but that doesn't mean SIMD can't make the application faster, it just takes some proper doing.
It took years for SIMD to become well utilized once C compilers added intrinsics exposing them to programmers. I expect a similar delay before userspace evolves to exploit io_uring, and we could really use better language-level constructs to make writing such async code more ergonomic.